TREE CROWN DELINEATION FROM DIGITAL ELEVATION MODELS AND
HIGH RESOLUTION IMAGERY
Christopher Meia , Sylvie Durrieub
a
INRIA, ICARE Project Team, 2004 Route des Lucioles - BP 93, 06902 Sophia Antipolis Cedex, France –
christopher.mei@sophia.inria.fr
b
UMR3S, Maison de la Télédection, 500 rue Jean-François Breton, 34093 Montpellier Cedex 5, France –
sylvie.durrieu@teledetection.fr
KEY WORDS :
forestry, high resolution multi-spectral imagery, infrared imagery, DEM/DTM, watershed algorithm, LIDAR, feature detection.
ABSTRACT :
The aim of this study was to develop a methodology to delineate accurately tree crowns using digital elevation models
independently of source (laser scanned data or stereoscopic pairs) and to identify the complementary information that can
be obtained from high resolution imagery. This methodology uses the Watershed algorithm as the baseline. However the
direct application of this algorithm generates over-segmentation for complex tree-crowns and non-wooded areas. To solve
these problems a preprocessing stage consisting of a mixture modeling adaptive thresholding technique was developed
to remove non-wooded areas. Results obtained on different plots of land are discussed putting forward the influential
factors.
1.
INTRODUCTION
There is an increasing need for accurate and cost effective
forest resource information for operational and strategic
applications (ecological sustainability, forest exploitation,
rural development). The combination of high resolution
digital elevation models, obtained for example from laser
scanned data, and multispectral images appears particularly promising as a source of forest survey tools (Leckie
et al., 2003).
Research in the area of high resolution imagery (HRI) has
led to the development of several methods for tree crown
delineation. Tree tops often appear as maxima in intensity in satellite pictures and cast shadows around the tree
crowns. In a 3D model where height corresponds to spectral intensity, tree crowns can be seen as peaks and shadows as valleys. Gougeon (Gougeon, 1998) developed a tree
crown delineation method using these properties with a
valley following approach. A similar technique was used
by (Warner et al., 1998) but this time using texture to
allow the detection of trees with a lower intensity. The results are generally good in high to medium density conifer
forests. Preprocessing to remove non-vegetated areas is
necessary for lower density areas.
Methods that use region approaches have also been studied
(Brandtberg & Walter, 1998; Culvenor et al., 1998). The
zero-crossings of the second derivative (Laplace operator)
can be used to detect parts of the tree crowns. A region
growing algorithm can then be used with the tree tops
(maxima spectral response) as seeds and light intensity
as growing function right up to the pre-calculated boundaries. These methods are more robust for detecting trees
which are partly in shadow. However preprocessing is also
needed to avoid detecting light patches in lower density areas. Pollock (Pollock, 1998) uses model matching, training
and spatial information to locate tree crowns. The advantage of this type of approach is that it avoids improbable
regions. However in Pollock’s method, the training is performed by an operator. This could be a disadvantage in
some circumstances. Voronoı̈ diagrams with a fuzzy approach have also been tested (Dubé et al., 1998). The use
of a probabilistic model is also an efficient way to avoid un-
likely regions. Research in the field of high resolution radar
sensors is recent. At first these were used as an efficient
way to generate digital terrain models (DTM) for orthorectification of digital satellite imagery. Later they were
used to study tree heights (Andersen et al., 2001) and to
identify trees by using the pulse response of the vegetation
(Pyysalo & Hyyppä, 2001). Numerous studies have confirmed the importance of small footprint LiDAR data for
forest inventories. Stand characteristics, like tree heights,
basal area and stand volume can be accurately estimated
by using laser scanning (Maltamo et al. (2004); Naesset
& Bjerknes (2001); Means et al. (2000)). Some countries
are considering in a close future the use of LiDAR technologies to perform their forest inventories (Nilsson et al.,
2003; Wulder, 2003).
Digital elevation models (DEM) obtained by LiDAR sensors or stereoscopic views have more rarely been used for
tree delineation. The chief advantage of DEMs is that they
are not affected by non-uniform light intensity. This means
that they can be used to detect trees that would otherwise
be in shadow. Obtaining tree heights is also an interesting
characteristic. However DEMs are often less precise than
HRI and trees that are very close cannot be extracted.
The aim of the work described in this article is to use the
complementary information from both a DEM and HRI of
the same region to obtain a robust tree crown delineation
algorithm. Although the DEMs employed in this study
have been obtained from LiDAR data and consequently
DTMs can also be extracted, it has been decided that the
approach should be DTM independent in order to be able
to use DEMs obtained from stereoscopic views.
2.
STUDY SITE AND DATA
The study site is a Mediterranean forest located near Montpellier city, in the south of France. In order to identify
the limits of the application of the method according to
the stands characteristics, six plots dissimilar in dominant
species, diameter-class distribution and tree spatial distribution were selected : a closed and an open mature
- 218 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
Aleppo pine (Pinus halepensis) stand, an umbrella pine
(Pinus pinea) young plantation, a mature poplar plantation and an olive orchard.
The data used in this study consists of :
• 1 m resolution digital elevation and terrain models
(DEM and DTM) produced by an engineering firm,
GeoLas Consulting, from LiDAR data with a 15 cm
horizontal precision and a 0.5 m vertical precision.
These data were acquired on 26th June 2002 with a
small-footprint Toposys laser scanning system.
DEM
HRI
• orthorectified multispectral images (blue, green, red
and Infra Red channels) with a 0.5 m resolution, acquired with a digital line camera that was operated
in parallel with the LiDAR instrument. All the data
are georeferenced to the RGF 93 coordinate system.
Pre−processed
DEM
Watershed
A ground truth database was available for the two Aleppo
pine stands and the Umbrella pine plantation. It includes,
for each plot, 1) individual tree location recorded in the
RGF 93 coordinate system and 2) dendrographic measurements (mean or both south-north and east-west crown diameters; total height, stem height, stem diameter at 1,30 m
above ground). For the poplar plantation and the olive orchard, trees were easily identifiable on the IRC images and
were manually delineated. The resulting vector layers were
considered as the ground reference.
3.
DTM
Pre−processing
Extraction
of vegetated
areas
Mask
Regions
HRI
Region
Improvement
METHOD
Tree crowns
3.1
The watershed algorithm
DEMs can be represented as grayscale images with the
pixel values representing altitude. By inverting the images, trees appear as catchment basins. These provide the
right configuration for the application of the Watershed
algorithm (Vincent & Soille, 1991).
Figure 1: Methodology for tree crown delineation
The results obtained by the direct application of the algorithm on a DEM led to 1) over-segmentation of most of
tree crowns and 2) imprecise borders obtained in low tree
density areas To overcome this problems a specific methodology was developed (Fig. 1).
3.2
Pre-processing the DEM
The over-segmentation of regions is not due to the classic
problem of the nature of the image (which would have led
to the application of a distance transform, a geodesic erosion or a marker-controlled segmentation). In this figure
the over-segmentation is due to the complex tree crowns
which do not conform to the basin assumption. Preprocessing with the application of a median or averaging
filter with a mask approaching the actual size of the trees
in the image can improve the results, the disadvantage being that some regions are then under-segmented.
3.3
Extracting wooded areas
The second problem to be solved is the imprecise borders
obtained for the tree crowns in low density areas. In this
case, the Watershed borders do not correspond to tree
crowns (Fig. 2).
Separating terrain and non terrain areas (including trees)
would solve this problem. Where a DTM is available, vegetated area extraction is possible by thresholding the image
difference DEM − DT M . This is not possible with the
DEM on non flat areas because the altitude of the terrain
True tree crown border
Tree crown border
found by the algorithm
True tree crown border
Tree crown border
found by the algorithm
Figure 2: Position of the watershed borders for high density wooded areas (left) and for low density wooded areas
(right)
- 219 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
is irregular.
• (µ1 − µ2 )2
A method has been developed to extract wooded areas
from DEMs (it was again decided to use DEMs instead of
HRIs to avoid the problem of shadowed trees). The histogram of a DEM on a small vegetated region frequently
contains two classes : the first corresponds to ground pixels, the second to tree tops. Spliting the DEM into squares
and estimating a threshold elevation value for each square
makes it possible to separate the trees from the ground
locally (Fig. 3).
• σ 1 , σ2
DEM
Image of thresholds
Gaussian blurring
Thresholded image
Incorrectly thresholded regions have been identified by comparing their variables to the Gaussian distribution generated by the variables of all the thresholded regions. This
approach can be seen as a way of learning the global distribution of trees over a given region.
A region can be incorrectly thresholded either because it
is homogeneous (only trees or only ground) or because the
trees are different to those in the rest of the DEM. By
increasing the size of the square used for the region, it may
be possible to make the region ressemble the surrounding
regions more closely and find a satisfactory threshold (Fig.
5).
Threshold
Figure 3: Adaptative thresholding for wooded areas extraction
This procedure is simplistic because some squares may not
contain two classes (both ground and trees). It is essential
to take into consideration the size of squares required to
obtain a robust separation of these two regions.
These difficulties have motivated the use of a Gaussian
gray-level histogram modelling algorithm where both
classes were modelled by a Gaussian that minimized the
approximation error (Fig. 4). The threshold is at the
intersection of both Gaussian functions. The functions’
characteristics can then be used to identify the regions
that have been incorrectly thresholded.
Figure 5: Result of the adaptive thresholding algorithm
before correction (left) and after correction (right)
3.4
After masking by applying the adaptative threshold, the
watershed algorithm is applied. Regions obtained in that
way may contain buildings. These non vegetated regions
can easily be identified using the normalized difference vegetation index (NDVI) on the HRI.
The higher precision of the HRI can also be used to improve borders of the identified tree crowns.
3.5
Max1
Analysis
Max2
Thresholding
0
mu1
mu2
Threshold
255
Altitude
Ground and tree heights are modeled
by two Gaussian functions
Figure 4: Mixture modelling applied to the histogram
The variables used to identify the incorrectly thresholded
regions must be independent of altitude (you can not just
compare the thresholds). The following variables were
used taking this into consideration :
•
•
max1
max2
, number of pixels
number of pixels per square
per square
(max1 − max2 )2
number of pixels per square
Automatic tree delineation
Validation
The results of the automatic terrain/non-terrain classifications are compared to a reference classification
by analysing the contingency matrices. The reference
classification was obtained by thresholding the Canopy
Height Model (CHM = DEM − DT M ). The threshold
was set manually so as to best suit the wooded areas
visible on the IRC image. Evolution of the classification
accuracy with the size of the square that was chosen to
split the DEM is analysed for the different stands and
related to the stand characteristics. The results were
analysed before the iterative improvement step.
The theoretical optimal size of the square used on the DEM
image in the terrain/non-terrain separation method can be
estimated for a regular stand. The window must contain
at least one pixel of each class (terrain/non-terrain) whatever its position. The optimal size of the window is then
a function of the tree diameters, distance between trees
and direction of the plantation. It varies quite a √
lot according to the stand characteristics, from d/2 to d 2 for
stands composed of trees at a distance d with perfectly
round crowns. For real stands we assumed that the mean
distance between trees for open stands, or between gaps
for closed stand, could give a rough approximation of the
optimal square size. This mean distance was estimated
- 220 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
from the density of trees (respectively gaps).
The results of the automatic tree delineation (called segmentation in what follows) are compared to ground reference delineation of trees. The delineation used as the
reference were obtained either by photo-interpretation of
the IRC images for uniform stands with well individualised trees or using a ground truth data base for the other
stands. A statistical analysis is first performed. The number and the size of automatically delineated trees are analysed for each selected plots and compared to the ones of
the reference delineation. A more precise analysis, which
can be called ”the spatial analysis of the segmentation”,
is performed by comparing the segment boundaries resulting from the automated recognition method to the ground
reference. The layers are compared in a vector format.
Seven types of overlaps are defined for segmentation accuracy assessment. A tree is well identified when there is
a one-to-one correspondence (only one segment associated
with one ground delineation and vice versa), with a tolerance of a two pixels gap between both boundaries and
with an overlap area greater than 80 % of the delineated
tree area. The other classes are : over-segmented trees
(more than one segment associated with one ground delineation) ; under-segmented trees (a segment includes
significant part (>10 %) of more than one tree); trees that
are both ”over-segmented” and ”under-segmented” that is
to say that among the several segments associated with
the ground delineation at least one is common to two or
more trees; trees that are well identified but with an overlap area <80 % of the delineated tree area ; omitted trees
; and finally segments not associated with a tree (commission errors).
4.
4.1
RESULTS
Extraction of wooded land
For five stands (olive orchard, poplar, dense pine stand,
sparse pine stand, young pine plantation) the evolution of
the classification accuracy with respect to the initial square
size for the extraction algorithm is analysed. For olive orchard, a large range of values was studied, from 1 to 40
meters with a one meter step. For other stands the studied size are spaced between 5 and 10 m. The maximum
classification rate and the corresponding window size are
related in table 1. The classification rate corresponding to
the square size which is the closest to the mean distance
between gaps or trees is also reported in table 1.
The maximum classification rates are high for all stands
(from 87.9 % to 98.7 %). For most of the stands the classification rates obtained with a square size close to the mean
distance between trees or gaps stay high except for poplar
and young pine plantations. The sharp decrease of the
classification accuracy rate (>10 %) for these stands can
be explained by the non homogeneous spatial distribution
of the gaps (all around the plot for example for the pine
plantation). In that case mean distance is no more a good
indicator of the optimal window size.
These results show the good potential of the proposed
method to separate terrain and non terrain. The adaptive step decreases the sensitivity to the initial size square
as expected.
4.2
Tree segmentation
The number of trees of the ground reference data and the
automatically delineated per radius classes are shown in
Fig. 8. The results of the spatial comparison between
the automated recognition method and the ground reference segmentation are presented in table 2. This spatial
analysis was only realised for two plots. For heterogeneous
stands (closed and open pine stands) the reference vector
layers obtained from the ground truth data base did not
fit enough the IRC images to be usable (tree crowns approximated with circles, shifts in tree locations).
The best results are obtained for the poplar plantation.
Only a slight over-segmentation is noticed (+6 % of the
trees). The mean radius of segments is greater than the
tree radius (+41 %) and this over estimation is observed
for all the radius classes (see the shift in the histograms on
fig $). However 80 % of the trees are well identified which
indicates that the over-estimation of the radius stays in
the limits set for an accurate segmentation. Taking into
account the class of ”well identified trees with a problem
of area” the number of well identified trees reaches 90 %.
For the olive orchard, an homogeneous stand with 71 %
of the trees in the same radius class, the results are not
as good. We can notice an over estimation of the number of trees obtained by automatic segmentation for all
the classes of radius except for the [2 m − 3 m[ radius class
that contains most of the trees. Trees were either over segmented or under segmented. This is finally expressed by a
number of segments greater than the real number of trees
(+26 %) and by a mean of segment radius approximately
the same as the real tree one (0.27 pixels). The well identified trees represents only 52.6 % of the trees (76.3 % if
well identified trees with a problem of area are included).
This result can be explained by the shape of olive trees.
These fruit trees are pruned in their centre in order to let
the light get into the middle of the trees. The crown shape
has no more the appearance of a catchment basin on the
inverted DEM but is closer to a torus which explains the
over-segmentation and misplacement.
For the two last stands analysed, the open pine stand and
the closed pine stand, most of the trees have a small diameter (class radius [ 0 - 1 [ m) and could not be identified
on the 1 m precision DEM. This led to the number of trees
being sharply under estimated for both stands (- 43.6 %
for open stand and -70 % for the closed stand).
5.
DISCUSSION
The terrain/non-terrain classification method gave
very encouraging results for all the studied stands.
Results concerning tree segmentation were much more
”stand-dependant”. The best results were obtained for
the poplar plantation with high and regularly spaced
trees (90 % of well identified trees). For natural pine
stands the segmentation gave poor results. This was not
surprising because of the complex shapes of the crowns
and the irregular spatial distribution of trees. For these
stands achieving an individual tree segmentation with the
available data is unlikely. However interesting information
concerning tree density, spatial organisation could be
derived from the automatic tree crown delineation. The
spatial resolution of the used DEM (1 m) allowed us
to identify trees with a radius equal to that resolution.
Better results are expected to be an outcome of DEM
improvement. Some promising tests to obtain a 0.5m
DEM from the same LiDAR data have already been
performed. More information could also be derived from
the complementarity of the DEM and HRI.
Future studies could be undertaken to improve the Water-
- 221 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
Stand
Olive trees
Poplar trees
Dense pine
Sparse pine
Young pine
Average
distance between
trees (or holes)
in m
7.8
9.5 (holes)
13.3 (holes)
18.2,
15.2(holes)
10.2 (holes)
Size of optimal
square in m
% of correctly
classified pixels
square closest
to average distance
% of correctly
classified pixels
11
25
20
10
91.8
98.7
87.9
84
8
10
10
20
90.6
86
85.7
83.8
30
88.2
10
68.8
Table 6: Accuracy of classification
Nb
Trees correctly segmented
Trees correctly segmented with a problem of area
Trees over-segmented
Trees under-segmented
Trees under- and over-segmented
Trees omitted
Extra segments
Total number of trees in the stand
Correct (segmentation+area)
20
4
5
4
5
0
0
38
Olive trees
%of total number of
trees of the stand
52.6
10.5
13.2
10.5
13.2
0
0
Nb
128
15
Poplars
%of total number of
trees of the stand
80.5
9.4
2
13
1
0
18
159
63.2
1.3
8.2
0.6
0
11.3
89.9
Table 7: Comparison between automated recognition and ground reference
Figure 8: Figures (a) to (d) show the numbers of trees per radius classes calculated for the ground reference data and
for the results of automatic delineation
- 222 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
shed algorithm using level sets. It would be interesting to
study automatic identification of optimal square size for
mixture modeling thresholding. Higher level knowledge
could lead to better results by identifying tree types. Extra information can easily be extracted from the regions
(tree crown diameter, height, ...) and could be used by
an identification algorithm (ex : neural network) to create
maps of regions automatically.
References
Andersen, H.-E., Reutebuch, S. E. & Schreuder, G. F.
(2001) Automated individual tree measurement through
morphological analysis of a lidar-based canopy surface
model. In 1st International Precision Forestry Symposium pp. 11–22, Seattle, Washington. University of
Washington, College of Forest Resources.
Brandtberg, T. & Walter, F. (1998) An algorithm for delineation of individual tree crowns in high spatial resolution aerial images using curved edge segments at multiple scales. In Automated Interpretation of High Spatial
Resolution Digital Imagery for Forestry pp. 41–54 Natural Resources Canada, Canadian Forest Service, Victoria, B.C., Canada.
Culvenor, D. S., Coops, N., Preston, R. & Tolhurst, K. G.
(1998) A spatial clustering approach to automated tree
crown delineation. In Automated Interpretation of High
Spatial Resolution Digital Imagery for Forestry pp. 67–
80 Natural Resources Canada, Canadian Forest Service,
Victoria, B.C., Canada.
H. U. (2003) Laser scanning of forest resources. ScandLaser Scientific Workshop on Airborne Laser scanning
of Forests., , 43–51.
Pollock, R. (1998) Individual tree recognition based on a
synthetic tree crown image model. In Automated Interpretation of High Spatial Resolution Digital Imagery for
Forestry pp. 25–34 Natural Resources Canada, Canadian Forest Service, Victoria, B.C., Canada.
Pyysalo, U. & Hyyppä, H. (2001) Reconstructing tree
crowns from laser scanner data for feature extraction.
In Automated Interpretation of High Spatial Resolution
Digital Imagery for Forestry Institute of Photogrammetry and Remote Sensing, Helsinki University of Technology.
Vincent, L. & Soille, P. (1991) Watersheds in digital
spaces: an efficient algorithm based on immersion simulations. IEEE PAMI, 1991, 13 (6), 583–598.
Warner, T. A., Lee, J. Y. & McGraw, J. B. (1998) Delineation and identification of individual trees in the
eastern deciduous forest. In Automated Interpretation
of High Spatial Resolution Digital Imagery for Forestry
pp. 81–91 Natural Resources Canada, Canadian Forest
Service, Victoria, B.C., Canada.
Wulder, M. A. (2003) The current status of laser scanning
of forests in canada and australia. ScandLaser Scientific
Workshop on Airborne Laser scanning of Forests, , 21–
33.
Dubé, P., Hay, G. J. & Marceau, D. J. (1998) Voronoı̈
diagrams, extended area stealing interpolation and tree
crown recognition : a fuzzy approach. In Automated Interpretation of High Spatial Resolution Digital Imagery
for Forestry pp. 115–125 Natural Resources Canada,
Canadian Forest Service, Victoria, B.C., Canada.
Gougeon, F. A. (1998) Automatic individual tree crown delineation using a valley-following algorithm and a rulebase system. In Automated Interpretation of High Spatial Resolution Digital Imagery for Forestry pp. 11–23
Natural Resources Canada, Canadian Forest Service,
Victoria, B.C., Canada.
Leckie, D., Gougeon, F., Hill, D., Quinn, R., Armstrong, L.
& Shreenan, R. (2003) Combined high-density lidar and
multispectral imagery for individual tree crown analysis.
Canadian Journal of Remote Sensing, 29, 633–649.
Maltamo, M., Eerikainen, K., Pitkanen, J., Hyyppä, J. &
Vehmas, M. (2004) Estimation of timber volume and
stem density based on scanning laser altimetry and expected tree size distribution functions. Remote Sensing
of Environment, 90, 319–330.
Means, J. E., Acker, S. A., Fitt, B. J., Renslow, M., Emerson, L. & Hendrix, C. J. (2000) Predicting forest stand
characteristics with airborne scanning lidar. Photogrammetric Engineering and Remote Sensing, 66, 1367–
1371.
Naesset, E. & Bjerknes, K.-O. (2001) Estimating tree
heights and number of stems in young forest stands using airborne laser scanner data. Remote Sensing of Environment, 78, 328–340.
Nilsson, M., Brandtberg, T., Hagner, O., Holmgren, J.,
Persson, A., Steinvall, O., Stemer, H. & Sderman, O.
- 223 -